Improving speaker diarization for naturalistic child-adult conversational interactions using contextual information

Area of Science:

• Speech processing
• Machine learning
• Human-computer interaction

Background:

• Deep learning advances speaker diarization but struggles with complex child-adult interactions.
• Varied acoustic environments and interaction dynamics pose significant challenges.

Purpose of the Study:

• To analyze contextual factors influencing speaker diarization performance in child-adult interactions.
• To identify specific factors contributing to diarization errors.
• To improve diarization accuracy by integrating contextual information.

Main Methods:

• Analysis of contextual factors affecting diarization errors in child-adult speech.
• Development and training of a deep neural network (DNN) incorporating these contextual factors.
• Evaluation of the DNN's performance on child-adult interaction data from clinical settings.

Main Results:

• Specific contextual factors were identified as key contributors to diarization errors.
• Training a DNN with contextual information led to enhanced diarization performance.
• Significant improvements in speaker diarization accuracy were observed for child-adult interactions.

Conclusions:

• Contextual factors play a crucial role in the accuracy of speaker diarization for child-adult interactions.
• Integrating contextual information into deep learning models is an effective strategy for improving diarization.
• This approach shows promise for applications in clinical settings and other challenging acoustic environments.